Agents Monitor

The Agents Monitor is the primary interface for watching AI agents work in real-time. It provides live WebSocket streaming, interactive terminals, and session management.

Overview

┌─────────────────────────────────────────────────┐
│  Agents Monitor                                  │
├─────────────┬───────────────────────────────────┤
│  Session    │  Live Stream                       │
│  List       │  ┌─────────────────────────────┐  │
│             │  │  Agent messages              │  │
│  • session1 │  │  Tool calls (Read/Edit/Bash) │  │
│  • session2 │  │  Thinking blocks             │  │
│  ◉ active   │  │  System events               │  │
│             │  └─────────────────────────────┘  │
│             │  ┌─────────────────────────────┐  │
│             │  │  Input bar                   │  │
│             │  └─────────────────────────────┘  │
└─────────────┴───────────────────────────────────┘

Features

Live Session Streaming

Sessions stream output over WebSocket connections. Each message type is visually distinct:

Agent text — formatted markdown with syntax highlighting
Tool calls — collapsible cards showing tool name, arguments, and results
Thinking blocks — italic, muted text showing internal reasoning
User messages — right-aligned bubbles
System events — centered dividers for session lifecycle events

Session Creation

Create sessions with configurable options:

Option	Values	Description
Agent Type	`claude`, `codex`, `openclaw`	Which AI agent to spawn
Permission Mode	`default`, `acceptEdits`, `bypassPermissions`	Tool approval behavior
Working Directory	Any path	Where the agent operates
Initial Prompt	Text	First instruction to the agent

Worker Dispatch

Sessions can dispatch worker sub-agents via the Factory module. Workers run in isolated git worktrees and report back to the parent session. The worker panel shows:

Active worker count and status
Per-worker task description
Link to worker session stream

Session Lifecycle

Created — session spawned, process starting
Active — agent running, streaming output
Idle — agent waiting for input
Completed — agent finished task
Killed — manually terminated

Debrief

Before killing a session, you can trigger a pre-kill debrief that generates either a hotwash (short sessions) or After-Action Review (long sessions). This captures what the agent accomplished and any issues encountered.

World View

The GET /api/agents/world endpoint provides a global view of all agents across all sessions, including their phase, role (commander vs worker), and resource usage.

Mobile Experience

On mobile devices, the Agents Monitor uses a full-screen overlay with:

Swipeable session list
Dark terminal theme with custom colors
Safe-area-aware input bar
Voice input via microphone button